Abstract:
|
There has been much development in nonparametric models to identify treatment effect heterogeneity in casual studies. While these methods have proven their ability to detect the presence of effect modification, they are often difficult to interpret and on their own do not describe significant interactions between the treatment and covariates of interest. Here, we propose a modular two-stage approach for creating parsimonious, interpretable summaries for models of treatment moderation. In the first stage, we fit a flexible model which is appropriately regularized to prevent inadvertent biased estimation of the treatment effect. In the second stage, we construct lower-dimensional summaries by projecting draws from the posterior of the conditional average treatment effect (CATE) function onto simpler structures. These summaries naturally come with Bayesian uncertainty estimates, and we use the the data only once to obtain the posterior for the response surface, thereby retaining the valid Bayesian uncertainty estimates across multiple summaries. We apply our techniques to an experiment measuring how teacher characteristics moderate the efficacy of an educational mindset intervention.
|